An RGB‐D Approach Assisted by Foundation Models for Reconstructing Bedforms With Artificial Vegetation

Detecting sand beds in laboratory sediment experiments is challenging, especially when artificial vegetation or other structures are present. RGB‐D cameras offer fast, high‐resolution, and low‐cost sand bed reconstruction. However, the commonly applied Simultaneous Localization and Mapping (SLAM)‐based approaches for RGB‐D data often fail on the flat and textureless sand surfaces with artificial structures, due to difficulties in RGB‐D‐based camera pose optimization. Additionally, accurate three‐dimensional point‐cloud segmentation remains difficult, and existing depth‐ or geometry‐based methods cannot reliably provide automatic separation of sand bed and vegetation point clouds. To address these challenges, we propose a hybrid workflow that transfers two‐dimensional (2‐D) masks from the Segment Anything Model to registered RGB‐D point clouds, enabling automatic three‐dimensional (3‐D) separation of sand‐bed and vegetation points. The workflow further applies threshold‐based filtering to correct edge gaps in depth maps. With predefined camera poses along a rail, we bypass SLAM and directly mosaic frames in physical coordinates, which improves reconstruction accuracy. This approach supports 3‐D bed reconstruction, bedform characterization (e.g., ripple shapes), and localized scour analysis behind vegetation stems. Across hundreds of monitoring points, we find low errors in scour depth and bed elevation, indicating that spatial variability is captured accurately. The workflow is straightforward and cost‐effective, lowering the barrier to high‐resolution sand bed reconstruction in laboratory flumes.